Figures on this page were verified 31 August 2026 against the providers' own documentation. Pricing, context windows and rate limits change without notice, so confirm any number against the provider before you rely on it. Tell us if something here is out of date.
Enter your token usage below and every major model is priced against the same workload, ranked cheapest first. On the default workload below (4,000 input tokens, 800 output tokens, 10,000 requests a month) the spread between the cheapest and most expensive model on this list is over 500x: about $5 a month against roughly $2,640. Output tokens, not input tokens, drive most of that difference.
Most cost calculators price one provider at a time, which tells you nothing about the decision you are actually making. This one puts Anthropic, OpenAI and Google side by side on identical inputs.
API cost calculator
API cost calculatorEnter one request’s token usage and your monthly volume. Every model is priced against the same workload and ranked cheapest first.
Runs entirely in your browser. Nothing is sent anywhere.
How API pricing actually works
Every provider bills two separate meters: tokens you send in, and tokens the model generates back. They are priced differently, and output is always the expensive one. Across the models in this table the output rate runs roughly four to eight times the input rate. This is the single most important fact about your bill.
A token is roughly three quarters of an English word. One million tokens is somewhere around 750,000 words on older tokenizers, though newer ones pack more text into the same count, so treat any word-to-token rule as an estimate rather than a conversion.
Why your bill is higher than your estimate
Four things routinely blow up a budget that looked fine on paper.
- Conversation history is resent every turn. In a multi-turn chat you are not paying for one message, you are paying for the whole transcript again on each request. A twenty-turn conversation can cost far more than twenty single requests.
- Retries are billed. A failed or malformed response still consumed tokens. Aggressive retry logic quietly multiplies cost.
- Reasoning tokens count as output. On models that think before answering, that thinking is billed at the output rate even when it is never shown to the user.
- Cache misses. Prompt caching only helps if the cached prefix stays byte-identical. A timestamp or a request ID near the front of your prompt invalidates everything after it.
The three levers that actually reduce cost
In rough order of how much they save per unit of effort:
- Shorten your outputs. Since output is priced several times higher than input, cutting generated length is the highest-leverage change available and it usually costs nothing in quality. Ask for less.
- Cache your stable prefix. Anthropic bills prompt cache reads at 10% of the base input rate. If your system prompt and tool definitions are large and unchanging, this is close to free money. Put volatile content last.
- Use batch processing where latency does not matter. Anthropic’s Batch API runs at 50% of standard pricing. If the work is not interactive, this is a straight halving.
Switching to a cheaper model is the obvious lever and usually the wrong one to reach for first. A cheaper model that needs two attempts to get the answer right is not cheaper. Judge cost per completed task, not cost per request.
A note on the numbers
Every price in this tool was read from the provider’s own documentation on the date stamped under the table, not copied from a secondary source. Providers change pricing without announcement, and some rates shown are explicitly promotional. Confirm against the official pages before you commit to a budget.
Batch and cache discounts are applied to Anthropic models only, because those multipliers are published as exact figures. Other providers offer comparable programmes, but hard-coding an unverified discount into a calculator would make it confidently wrong, which is worse than incomplete.
Frequently asked questions
Does this calculator send my data anywhere?
No. It runs entirely in your browser. Nothing you type is uploaded, stored or logged, and the page makes no network calls once it has loaded.
Why is my real bill higher than the estimate?
The usual causes are conversation history being resent on every turn, retries after failures, cache misses, reasoning tokens on reasoning models, and output tokens costing several times more per token than input. This tool prices one call pattern; a live workload multiplies it.
How current is the pricing?
Every figure carries the date it was verified against the provider's published pricing, shown at the top of this page. Providers change prices without notice, so confirm against their pricing page before you commit to a budget.